ARIA is an AI assistant that lives in your terminal. It knows about Termux, Android filesystem quirks, proot containers, clang-not-gcc, and all the weird stuff you deal with when developing on a phone.
It can answer questions, diagnose errors, suggest fixes, and auto-repair broken installs — all from a slash-command interface inside Termux.
It's lightweight. One dependency (rich). Everything else is stdlib. No bloated frameworks.
VID_20260529_060131_850.mp4
pkg update && pkg upgrade -y
pkg install python git -y
git clone https://github.com/Alex72-py/aria-termux.git
cd aria-termux
pip install -r requirements.txt --break-system-packages
python run_aria.pyThat's it. If you want to use the Google provider specifically, also run:
pip install google-generativeai --break-system-packagesOpenRouter and NVIDIA NIM work out of the box — no extra packages.
On first launch, ARIA detects there's no config and walks you through setup:
Choose provider [google/openrouter/nvidia_nim]: google
Enter API key: ••••••••••••••
Enter model [gemma-4-26b-a4b-it]:
Enable Guardian mode? [y/n]: y
It validates your key right there. If something's wrong, you'll know immediately instead of getting cryptic errors later.
Get a free key:
- Google: aistudio.google.com/app/apikey
- OpenRouter: openrouter.ai/keys
- NVIDIA NIM: build.nvidia.com
Just type. Bare text goes to the model as a question. Or use slash commands:
| Command | What it does |
|---|---|
/ask <question> |
Ask anything |
/fix |
Diagnose the last failed command |
/fix "error text" |
Diagnose a specific error |
/provider list |
Show configured providers |
/provider openrouter |
Switch provider |
/provider cycle |
Rotate to next provider |
/provider key <name> |
Update a provider's API key |
/model list |
List available models |
/model set <name> |
Switch model |
/kb <query> |
Search offline knowledge base |
/watch |
Toggle watch mode (monitors errors from other sessions) |
/status |
Show current state |
/config |
Re-run the setup wizard |
/help |
Show all commands |
Enable /watch, then work in another Termux session. When a command fails there, the shell hook writes it to ~/.aria/last_fail.json. Come back to ARIA, type /fix, and it reads the failure automatically.
| Provider | Endpoint | Auth |
|---|---|---|
google |
Google AI Studio | API key from aistudio |
openrouter |
openrouter.ai | API key from dashboard |
nvidia_nim |
integrate.api.nvidia.com | API key from build.nvidia.com |
ARIA auto-discovers available models from whichever provider you're using.
Gemma 4 family (Google):
├── gemma-4-2b-it
├── gemma-4-4b-it
├── gemma-4-26b-a4b-it
└── gemma-4-31b-it
aria/
├── main.py — Main loop, command dispatch, setup wizard
├── api_client.py — Provider abstraction (Google, OpenRouter, NVIDIA)
├── command_system.py — Slash command registry and parsing
├── config.py — JSON config management (~/.aria/config.json)
├── guardian.py — Risk scoring and confirmation prompts
├── knowledge_base.py — Offline Termux troubleshooting data
├── watch_mode.py — File watcher for cross-session error capture
├── repair_agent.py — Local-first fix planning (no model call needed)
└── ui.py — Rich terminal output, spinners, formatting
The API layer uses Python's urllib directly — no requests, no httpx, no heavy HTTP libs. Provider switching is just changing which base URL and auth header gets used.
Before ARIA suggests running anything dangerous, it scores the risk:
| Level | What happens |
|---|---|
| Low (green) | Runs immediately |
| Medium (yellow) | Runs with a notice |
| High (orange) | Asks for confirmation |
| Critical (red) | Requires explicit approval |
Things like rm -rf, chmod 777, pipe-to-shell curls — all get flagged.
Stored at ~/.aria/config.json:
{
"provider": "google",
"api_key": "your-active-key",
"api_keys": {
"google": "...",
"openrouter": "...",
"nvidia_nim": "..."
},
"model": "gemma-4-26b-a4b-it",
"temperature": 0.7,
"max_tokens": 8192,
"guardian_mode": true,
"watch_mode": false,
"auto_apply": false
}You can also override with env vars:
export ARIA_API_KEY="..."
export ARIA_PROVIDER="openrouter"
export ARIA_MODEL="gemma-4-31b-it"| Package | Why | Required |
|---|---|---|
rich |
Terminal UI, syntax highlighting, spinners | Yes |
google-generativeai |
Google provider only | No |
Everything else — HTTP, JSON, threading, config — is stdlib. OpenRouter and NVIDIA NIM need zero extra packages.
pytest -q # 34 tests, all pass
pytest -v # verbose- Fixed the stuck spinner (model responses now terminate the animation properly)
- First-run auto-detects missing config and launches setup wizard
- API keys get validated on save (real auth check, not just endpoint ping)
- Removed streaming mode — non-stream is more reliable on mobile connections
google-generativeaiis now optional, not required- Fixed
install.shhardcoding wrong directory name - Config schema aligned between installer and runtime
- Watch mode works via file polling — it's not instant, but it's reliable
- Needs internet for AI features (knowledge base works offline)
- Built for Termux on Android — works on Linux too, but that's not the focus
- Auto-fix suggestions should be reviewed before applying
git checkout -b feature/my-thing
# make changes
git commit -m "what you did"
git push origin feature/my-thingKeep it simple. PEP 8. Tests for new features. Don't add dependencies unless absolutely necessary.
MIT. See LICENSE.
Submitted for the Google Gemma 4 Challenge
